AI Escape Incident Tracker (ai-escape.watch)

The newly launched AI Escape Incident Tracker presents a comprehensive database that chronicles failures in AI control systems, focusing specifically on the actions of agents when they breach their defined parameters. Unlike existing databases that catalog instances of AI-related harms, this tracker emphasizes what went wrong in terms of the controls that failed to prevent these incidents. Each entry is strategically organized by occurrence month and contains a Containment Breach Score (CBS) that assesses how far the AI advanced unchecked, based on a set of weighted factors. Significantly, this tracker allows for better identification of common vulnerabilities in AI controls, highlighting which aspects require immediate attention and funding. By linking incidents to specific ineffective guardrails, the project aims to construct a clearer pattern of failures that can inform future AI governance and risk management strategies. The inclusion of a draft phase invites community feedback and encourages transparency in the ongoing discourse surrounding AI safety, ensuring that the registry remains a dynamic tool for understanding and mitigating AI-related risks.